[XviD-devel] R-D "optimal"

[Marco Al]

From: "Christoph Lampert" <chl at math.uni-bonn.de>

> What can we do about this? Do you know about how to solve this?

Let the user set a lambda? (Which you then use for ME and quantization
alike.)

In 2 pass you could determine an approximate lambda needed for the final
filesize, reasoning from say p-domain statistics (possibly for several
different classes of MB depending on variance as in the thesis). Should make
XviD do really well in MSE/PSNR comparisons (very much questionable wether
this will be reflected in quality though).

A little closer to optimal, in a R-D sense with distortion being measured by
MSE/PSNR, still far from of course.

It still doesnt deal with the interdependence of MV and texture bits, the
lack of optimal DQUANTs (requires another trellis search), the problems of
pure scan-pattern search ME (the problem of the present MV's effect on rate
of subsequent MVs is ignored), the total intractibility of determining an
optimal b-frame sequence ... and last but not least the fact that R-D
optimal frame compression does not give R-D optimal sequence compression
even with constant lambda (the interdependence between subsequent frames is
rather significant, probably moreso than for motion and texture bits).

:)

Marco